Map to Map: From SLAM to CAD Maps and Back Using Generative Models

This is the code implementation of the paper From SLAM to CAD Maps and Back Using Generative Models. The code relies on the Anime2clothing repository which is included as part of this code.

Flowchart of proposed system is shown with the SLAM-to-CAD application of Section III and the CAD-to-SLAM application of Section IV in the paper:

Preparations and Training Options

• create invironment:

conda env create -f environment.yml
conda activate rema

• Unzip the dataset file
• To include cropping in the training use dataset_cropping.py instead of dataset.py
• To decide on how long to keep the training, change in train_options.py the epoch mapping giving the amount of epochs required at every resolution
• When running train.py set niter and niter_decay to 70% and 30 % respectively of the total number of epochs.

For SLAM to CAD (BtoA):

• Epoch mapping: Use 32 epoch mapping (400000 epochs for 32 resolution)
• No cropping (use dataset.py with no cropping)

python train.py --project_name BA --dataset DATASET_proc_nopartial_BA_thick --niter 525 --niter_decay 225

For CAD to SLAM (AtoB):

• Restore epoch mapping to how it was (15,30...)
• Cropping (use dataset.py with cropping)

python train.py --project_name AB --dataset DATASET_proc_nopartial --niter 140 --niter_decay 60

Testing

• As the model is training, it will test on the testing data.
• The testing file is taken as is from Anime2clothing and thus will not give you the same results generated using the train.py.
• Some pretrained models can be found here.

Reference

If you use parts of this code please cite this work:

@phdthesis{daher2020map,
  title={Map to Map: From SLAM to CAD Maps and Back Using Generative Models},
  author={Daher, Rema},
  year={2020}
}